Uso de espectrofotometria FTIR (Fourier Transform Infrared) e mineração de dados para a detecção e identificação de adulterantes no leite cru

Abstract

Milk is a high biological value food often involved in fraud, whose practice generates not only economic losses, but also risks to consumer health. Currently, the methods for detecting adulterants in raw milk provided by Brazilian legislation have limitations in relation to their analytical sensitivity, as well as being time consuming, consuming large quantities of reagents and generating pollutant residues. For these reasons, they have been replaced by more efficient instrumentation methods, such as FTIR spectroscopy, associated with data mining techniques that allow the detection and identification of these adulterants. The objective of this work was to detect and identify the adulterations in the spectral data of the milk samples analyzed by the FTIR spectrophotometer, through the classifications by deep and ensemble learning. A total of 9,788 milk samples were evaluated, of which 2,376 were adulterated with starch, sucrose, sodium bicarbonate, hydrogen peroxide and formaldehyde at different concentrations, temperatures and storage times. Different classifiers were used to train models capable of recognizing the alterations caused by the adulterants in the characteristics of normal milk composition. Binary and multiclass classifications were performed with the selected training and test subsets for the Gradient Boosting Machine (GBM), Random Forests (RF) and Convolutional Neural Networks (CNN) classifiers. The classification was performed using two types of data: the total infrared spectrum was analyzed by CNN, and the numerical components extracted from the equipment, by GBM and RF classifiers. For the ensemble methods (GBM and RF), the classification accuracies ranged from 93.18% to 98.72%. The CNN proposal, however, produced precision of up to 99.34%. Both methods presented high precision, but the CNN obtained better results, since it uses a more dense set of data (spectral coordinates). In other words, according to the proposed CNN architecture, one can predict with >99% accuracy that the analyzed sample is unadulterated (screening method) and, even more so, to identify which adulterant is added in the trained model, greatly contributing to the agricultural inspection, aiming at the guarantee of authenticity, quality and public health.